Complex CNN on complex images gives RuntimeError

complex
1

Hi,

I am trying to train a UNet model on complex images (MRI data), and, while I can convert the model and get the complex images (both cfloat, same happens with cdouble) just fine, I get the following error during the forward pass:

RuntimeError: getCudnnDataTypeFromScalarType() not supported for ComplexDouble

And the full stack trace:

Traceback (most recent call last):
  File "<MY FOLDER>/train.py", line 105, in <module>
    main()
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/hydra/main.py", line 32, in decorated_main
    _run_hydra(
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/hydra/_internal/utils.py", line 344, in _run_hydra
    run_and_report(
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/hydra/_internal/utils.py", line 199, in run_and_report
    raise ex
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/hydra/_internal/utils.py", line 196, in run_and_report
    return func()
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/hydra/_internal/utils.py", line 345, in <lambda>
    lambda: hydra.run(
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/hydra/_internal/hydra.py", line 107, in run
    return run_job(
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/hydra/core/utils.py", line 126, in run_job
    ret.return_value = task_function(task_cfg)
  File "<MY FOLDER>/train.py", line 98, in main
    start_single_process(0, n_gpu, working_dir, config)
  File "<MY FOLDER>/train.py", line 83, in start_single_process
    train_worker(config)
  File "<MY FOLDER>/train.py", line 46, in train_worker
    trainer.train()
  File "<MY FOLDER>/srcs/trainer/base.py", line 81, in train
    result = self._train_epoch(epoch)
  File "<MY FOLDER>/srcs/trainer/trainer.py", line 52, in _train_epoch
    output = self.model(data)
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "<MY FOLDER>/srcs/model/unet/unet.py", line 24, in forward
    x1 = self.inc(x)
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "<MY FOLDER>/srcs/model/unet/unet_parts.py", line 25, in forward
    return self.double_conv(x)
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/torch/nn/modules/container.py", line 117, in forward
    input = module(input)
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 423, in forward
    return self._conv_forward(input, self.weight)
  File "<PYTHON FOLDER>/miniconda3/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 419, in _conv_forward
    return F.conv2d(input, weight, self.bias, self.stride,
RuntimeError: getCudnnDataTypeFromScalarType() not supported for ComplexDouble

I wasn’t although able to pinpoint the cause. Has anyone come across the same issue before?

2

Cudnn doesn’t support the ComplexDouble data type, so you would have to disable it via torch.backends.cudnn.enabled = False and rerun the code.

3

Thank you very much for the reply. Doing that causes another similar error

RuntimeError: _thnn_conv2d_forward not supported on CUDAType for ComplexFloat

For the same operation (F.conv2d)

4

Thanks for the update. In that case neither the cudnn nor the native conv kernels support complex inputs at the moment. As a workaround you could use tensor.unfold (same as the im2col operation) and apply a matrix multiplication using the complex inputs/fiters.

5

I am getting this error while running a downloaded program which i am using for my understanding. Input values are in the form of complex number but i am not be able to initialize the weights for convolutional autoencoder in complex number form. so i am getting this error anyone please help me how to initialize weights in complex number form.
Error message:
‘’‘Input type (CUDAComplexDoubleType) and weight type (torch.cuda.FloatTensor) should be the same’’’

6

Hey Patrick, Does this command work in torch 1.9.0?

Says: Cannot find reference ‘cudnn’ in ‘init.py’

7

I’m not sure which command you are referring to, but if it’s torch.backends.cudnn.enabled, then yes it’s still working in 1.9.0.